An analysis of neural models for walking control

Area of Science:

• Computational Neuroscience
• Robotics
• Evolutionary Computation

Background:

• Numerous neural models exist, varying in complexity and biological relevance.
• Prior research focused on biological realism or mathematical tractability, not expressive power for control tasks.
• Sensorimotor control, especially locomotion, requires sophisticated neural controllers.

Purpose of the Study:

• To compare the expressive power of neural models with varying complexity for sensorimotor control.
• To determine if more sophisticated neural models outperform simpler ones in complex control tasks.
• To investigate the collective problem-solving capabilities of simpler neural networks.

Main Methods:

• Evolved neural network controllers using a genetic algorithm to achieve locomotion in a simulated four-legged robot.
• Selected four neural models with diverse complexity levels, ranging from simple abstractions to biophysical representations.
• Utilized a dynamically stable robot simulation with tight, time-dependent sensorimotor coupling.

Main Results:

• The most complex, biologically based neural model demonstrated significantly superior performance in walking control.
• The advanced model successfully generated recognizable gaits, indicating effective sensorimotor control.
• Simpler neural models, despite their collective potential, were less effective in solving this complex task.

Conclusions:

• Biologically detailed neural models are more effective for complex sensorimotor control tasks.
• Model complexity and biological realism are crucial factors for achieving robust locomotion control.
• This study highlights the advantage of sophisticated neural architectures in dynamic control challenges.